Add BBQ xLLM support

[akhauriyash]

Summary

• Add xLLM 375B and 8B model-args scripts for BBQ runs.
• Register xLLM mbridge and Megatron-to-HF conversion support.
• Keep long-context PPO logits in model precision until chunked logprob computation to avoid full-vocab fp32 materialization OOMs.
• Clear inactive CUDA allocator blocks before optimizer step and checkpoint saves for large packed RL batches.

Notes

• The 8B model args use --layernorm-num-groups 4, which depends on the corresponding Megatron-LM grouped RMSNorm support.
• This branch was created from the RL360-pinned Miles commit 22d4693bf.

Validation

• python3 -m py_compile on touched Python files.
• bash -n on scripts/models/xllm-375B.sh and scripts/models/xllm-8B.sh.
• git diff --cached --check before commit.

[nightlessbaron]

Tested this PR e2e on 8B BBQ.